EquiPulse AI

Millions of micro-businesses in emerging markets run their local economies but face substantial revenue losses due to an inability to predict demand, handle mathematical reconciliation, or access real-time retail intelligence. Existing ERP and POS networks are expensive, internet-dependent, and built for massive corporations rather than the small merchants of developing nations like Bangladesh. EquiPulse AI bridges this gap with a fully localized, offline-first business intelligence system engineered specifically for local merchants. Built with a simple, swipe-based user interface, the system stays operational even when internet access drops, shifting to local on-device fallback AI model pipelines without stalling business transactions. Technical Core and AI Architecture: 1. Intelligent Data Analytics: A user uploads a standard transaction CSV into the system. Our browser-based client-side DuckDB engine analyzes transactional data instantly, dynamically updating UI widgets and analytics components via a global POS context layer. 2. Cross-Framework Action Cards: Powered by Gemini 1.5 Flash inference, the system builds structured action cards for stock restocking, optimal price adjustments, and demand forecasting. These trigger instant automated workflows via webhooks. 3. Multimodal Receipt Parsing: A localized multimodal OCR engine scans physical Bangla invoices, automatically maps inventory relationships, and handles mathematical reconciliation. 4. Hybrid Graph RAG Layer: Instead of feeding huge unstructured logs into the models, our symmetric scrolling engine queries a custom hybrid Graph RAG pipeline. This captures intricate localized behaviors like massive seasonal demand shifts (e.g., Eid or Pohela Boishakh) for high-accuracy local inferences. EquiPulse AI operates on a tech stack combining Gemini AI, local Llama models, DuckDB, IndexedDB, and Firebase security rules. Furthermore, moving from a smart concept to a production-ready engine.